If a page displays images via data: uris, some form of memory usage related to the data: uris is never released. If javascript is used to dynamically create image elements on page (via createElement, innerHTML etc) using data: uris as the source, memory usage will increase and will not decrease proportionally when the image is subsequently removed from the page. There is a brief writeup of this bug at http://waldheinz.de/2010/06/webkit-leaks-data-uris/ , and a reduced test case based on that page at http://turmeric.assanka.com/rowan/imgmemory.html . I initially thought this might be related to the page image cache, but that appears to be capping out correctly (at ~128MB IIRC?) if images loaded via URL instead of data URI are used instead. That also suggests this isn't related to some of the other rapid-loading image issues ("death by a thousand papercuts" etc), and instead may be related to a URL issue such as https://bugs.webkit.org/show_bug.cgi?id=19699 . There's a Chromium bug open for this, although that bug seems to cover several unrelated image issues; jamesr has done some investigative work, see his comment at http://code.google.com/p/chromium/issues/detail?id=36142#c64 . This causes severe leaks on latest nightlies, Safari, and Chrome, and crashes on iOS.
We probably have a duplicate for this already, but as you've mentioned, it's hard to tell since there are usually several issues conflated in each bug report.